Configure the Smart Token Engine
Open Settings → Smart Token Engine to change how Kora prepares context and generation parameters for future messages.
Choose an optimization mode
Section titled “Choose an optimization mode”| Mode | App behavior |
|---|---|
| Auto | Selects Eco, Balanced, or Quality from the next turn’s model context window |
| Eco | Uses lower generation variability |
| Balanced | Uses the default generation settings |
| Quality | Uses higher generation variability and disables Output Optimization |
The mode has no effect when every optimization layer is off. Use the optimization control in the composer, next to the execution-mode selector, to override the global mode for the current session; its dropdown lists every mode with a one-line description.
Inspect context usage
Section titled “Inspect context usage”The context indicator in the chat shows the latest measured or estimated occupancy against the active model’s context window. Select it to inspect:
- total tokens used and the percentage of the context window consumed;
- the relative contribution of messages, system tools, skills, system prompt, MCP tools, and metadata;
- the weighted cache hit rate accumulated across completed requests for the same provider and model.
The breakdown is based on the exact prepared request when available. Provider-reported totals remain authoritative; after compaction, the previous breakdown is invalidated until a new request is completed.
When the context safety limit blocks a request, the error names the way out: compact the session from the context indicator and resend the message.
Use Auto with the active model
Section titled “Use Auto with the active model”Select Auto as the first option in Settings → Smart Token Engine when the default should follow the model selected for a turn:
| Model context window | Effective mode |
|---|---|
Up to 128k |
Eco |
More than 128k and up to 512k |
Balanced |
More than 512k |
Quality |
The composer control shows the result as Auto · Eco, Auto · Balanced, or Auto · Quality. Changing models between turns recalculates that result before the next request. The selection uses the model’s context-window capacity, not the tokens already used in the conversation.
The composer control is a session preference. You can switch between Auto and a manual mode any number of times in the same session; each change applies to the next message. It overrides the global selection for that session; new sessions start from the global selection in Settings.
Enable optimization layers
Section titled “Enable optimization layers”| Control | Effect in the app |
|---|---|
| Adaptive Parameters | Applies the generation parameters for the selected mode |
| Output Optimization | Instructs the model to keep output concise; unavailable in Quality mode |
| Predictive Compaction | Compacts the conversation before it reaches the model context limit |
| Adaptive Context | Selects older messages by relevance while retaining recent messages |
Set the global temperature
Section titled “Set the global temperature”Use Global temperature override to replace the mode temperature for every model. The range is 0.0–2.0. Click Reset to return to the selected mode. Temperature precedence is: selected STE profile, global override, then per-model temperature. The per-model setting wins.
Set the compaction ceiling
Section titled “Set the compaction ceiling”When Predictive Compaction is enabled, set Autocompaction ceiling between 50% and 95%. Kora starts proactive compaction five percentage points before the selected ceiling. Click Reset to restore 90%.
Configure Adaptive Context
Section titled “Configure Adaptive Context”When Adaptive Context is enabled:
- Top-K Messages sets the maximum number of messages selected by relevance.
- Always Keep Recent sets how many recent messages remain in context regardless of relevance.
Use a bare prompt with local models
Section titled “Use a bare prompt with local models”Enable Local model bare prompt to remove Kora-generated system context for custom providers on localhost. Leave the optional prompt empty to send no system field, or enter the minimal system prompt to send. Remote providers ignore this setting.
Apply a configuration
Section titled “Apply a configuration”- Open Settings → Smart Token Engine.
- Select Auto, Eco, Balanced, or Quality.
- Enable the layers you want Kora to apply.
- Adjust the controls shown for those layers.
- Close Settings. The changes apply to the next message.
Troubleshooting
Section titled “Troubleshooting”- The selected mode has no effect → enable at least one optimization layer.
- Responses are too concise → disable Output Optimization or select Quality.
- Conversation history is compacted too early → raise Autocompaction ceiling.
- A local model still receives Kora context → confirm its base URL uses localhost and Local model bare prompt is enabled.